OpenAI topped the leaderboard with a perfect score at the ICPC 2025 programming competition, and Gemini also achieved a gold medal level.

After the IMO, both OpenAI and Gemini won the gold medals at the ICPC 2025.

Just now, both OpenAI and Gemini claimed to have reached the gold - medal level of the ICPC.

Among them, OpenAI solved all 12 problems within 5 hours, equivalent to ranking first among humans and surpassing all participating university teams.

Meanwhile, Gemini solved 10 out of 12 problems in a total of 677 minutes, achieving a gold - medal level. If compared with human teams, it would rank second.

In terms of human teams, the team from Saint Petersburg State University in Russia ranked first, solving 11 problems. The teams from Beijing Jiaotong University, Tsinghua University, Peking University, and the University of Science and Technology of China ranked 2nd, 4th, 5th, and 9th respectively.

 The ICPC, or the International Collegiate Programming Contest, is the world's oldest, largest, and most prestigious university - level algorithm programming competition. It is one level higher than high - school Olympiads such as the IMO. Every year, participants from nearly 3,000 universities and 103 countries gather to tackle real - world programming challenges.

This year's ICPC World Finals were held in Baku, Azerbaijan on September 4, bringing together the top teams from the earlier stages of the competition. During the five - hour competition, each team solved a set of complex algorithmic problems. The final ranking was strictly based on two principles: only perfect solutions could score points, and every minute counted. Among the 139 participating teams, only the top four teams won gold medals.

Below are the original ICPC problems. Interested readers can try them out for themselves.

https://worldfinals.icpc.global/problems/2025/finals/index.html

OpenAI Solved 12 Problems in 5 Hours, Surpassing Human Teams

OpenAI competed under exactly the same conditions as the world's top human contestants: facing the same set of competition questions, having the same 5 - hour time limit, and being judged in real - time by a local system consistent with the standards of the ICPC Global Finals.

Throughout the process, the AI system independently analyzed the problems and decided to submit the final answers without the assistance of any customized testing tools.

The competition results were remarkable: out of all 12 problems, the AI system got the correct answers on its first submission for 11 of them. Even for the last problem, which was the most difficult in the whole competition and stumped all human teams, the AI managed to solve it after 9 attempts. In contrast, the best - performing human team in this competition successfully solved 11 problems.

For Problem G, OpenAI solved it after 9 attempts. This problem was also one of the two difficult problems that DeepMind failed to solve. For reference, the fastest - solving human contestant also took 270 minutes (the total competition time was 300 minutes).

OpenAI revealed that the AI participating in this competition was composed of a "general reasoning model ensemble" and was not specifically optimized or trained for the ICPC.

During the problem - solving process, the system combined its next - generation model GPT - 5 with a cutting - edge experimental reasoning model. Among them, GPT - 5 accurately solved 11 problems, while the experimental model finally made the crucial breakthrough for the most difficult problem.

This achievement is an excellent milestone in OpenAI's series of demonstrations of the astonishing progress rate of its reasoning systems. The same set of models has proven its strength in competitions such as the International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI), fully confirming its strong generality and wide - ranging application potential.

OpenAI employees Borys Minaiev and Mostafa Rohaninejad also posted on X to celebrate.

Borys Minaiev

Borys Minaiev is a researcher at OpenAI, focusing on the development and application of large - scale reasoning models, and has demonstrated outstanding abilities especially in programming competitions and complex reasoning tasks.

He graduated from Saint Petersburg State University of Information Technologies, Mechanics and Optics (ITMO University) and has achieved remarkable results in the field of programming competitions. In 2015, as a member of the ITMO University team, he won the championship at the International Collegiate Programming Contest (ICPC) World Finals. This was the only team in the history of the competition to solve all problems before the end of the competition.

After joining OpenAI, Borys Minaiev became one of the core members of the large - scale reasoning model research, participating in several key projects, including the development of models such as o1, o3, and o4 - mini.

In addition, Borys Minaiev is also active in the open - source community, sharing several projects on GitHub and delving into topics such as simulated annealing algorithms, the Rust programming language, and the application of AI in education on his personal blog.

Mostafa Rohaninejad

Mostafa Rohaninejad is a research scientist at OpenAI, focusing on meta - learning, reinforcement learning, and the reasoning ability of artificial intelligence systems.

He joined OpenAI in 2023 and participated in several key projects, including the development of large - scale reasoning models such as GPT - 5 and OpenAI o1.

Before joining OpenAI, Mostafa pursued a master's degree in computer science at the University of California, Berkeley, and collaborated with Professor Pieter Abbeel at the BAIR Laboratory of the university to research meta - learning and generative models. He is a co - author of the famous SNAIL architecture, which performs well in few - shot learning tasks.

Mostafa's research interests mainly focus on how to enable artificial intelligence systems to have stronger reasoning abilities and adaptability, especially in complex tasks and dynamic environments. His work at OpenAI not only promotes the development of AI technology but also lays the foundation for realizing more intelligent and user - friendly AI systems.

Google's Gemini Solved 10 Difficult Problems, Reaching the Gold - Medal Level

The advanced version of Gemini 2.5 Deep Think participated in the competition in a remote online environment under the ICPC rules and was conducted under the guidance of the competition organizers.

It started 10 minutes later than human contestants but correctly solved 10 out of 12 problems within the five - hour time limit, achieving a gold - medal - level performance.

Gemini 2025 ICPC World Finals code: https://github.com/google - deepmind/gemini_icpc2025

Gemini solved 8 problems in just 45 minutes and then solved two more problems in three hours, using various advanced data structures and algorithms to generate solutions. By solving 10 problems in a total of 677 minutes, if compared with the results of university teams, Gemini 2.5 Deep Think would rank second.

The picture shows the problem - solving time for each problem in the 2025 ICPC World Finals. Gemini's times are shown in blue, and the times of the fastest university teams are shown in gray.

Notably, Gemini successfully solved Problem C within half an hour, while no university team solved this problem in the competition.

This problem required finding a solution to distribute liquid to multiple reservoirs through a series of interconnected pipes, with the goal of finding a configuration that would fill all reservoirs as quickly as possible. Since each pipe could be open, closed, or even partially open, there were an infinite number of possible configurations, making it very difficult to find the optimal configuration.

Gemini found an effective solution: it first assumed that each reservoir had a "priority value", representing the preference of that reservoir relative to other reservoirs. Given a set of priority values, the optimal pipe configuration could be found through a dynamic programming algorithm. Gemini discovered that by applying the minimax theorem, the original problem could be transformed into finding the priority value that most restricted the flow. Using the relationship between the priority values and the optimal flow, Gemini quickly found the optimal priority value through nested ternary searches, thus successfully solving Problem C.

According to Google's internal research, a similar version of Gemini 2.5 Deep Think could also achieve gold - medal - level performance in the 2023 and 2024 ICPC World Finals, comparable to the performance of the world's top 20 programming contestants.

In addition, Google's official blog also thanked the many contributors behind this project. Among them, Hanzhao (Maggie) Lin led the overall technical direction of Gemini's competitive programming and the ICPC 2025 work, and co - led the overall research and execution work with Heng - Tze Cheng.

Hanzhao (Maggie) Lin

Hanzhao (Maggie) Lin is a senior research scientist at Google DeepMind, focusing on the research and development of large - scale language models and multimodal systems.

Her research directions mainly cover large - scale language models, system architectures, and their applications in education and complex reasoning. Her contributions in the field of AI include participating in the post - training research of large - scale language models such as LaMDA and PaLM 2 at Google DeepMind and promoting the improvement of the models' abilities in multimodal understanding, reasoning, and tool use.

In addition, she also led the application of the Gemini Deep Think model in the International Mathematical Olympiad (IMO) competition, achieving gold - medal - level performance and demonstrating the potential of AI in complex mathematical reasoning.

Heng - Tze Cheng

Heng - Tze Cheng is the research director and chief research scientist at Google DeepMind, focusing on the research and application of large - language models and conversational AI. He has a deep research background in fields such as natural language processing (NLP), recommendation systems, reinforcement learning, and multimodal reasoning.

He graduated from the Department of Electrical Engineering at National Taiwan University with a bachelor's degree and obtained a doctorate in electrical and computer engineering from Carnegie Mellon University in 2013, with research